A Knowledge-Guided Graph Reasoning Framework for Multi-Modal Detection of Mining-Induced Land Degradation

The unprecedented explosion of multi-modal geospatial data offers a transformative opportunity for environmental surveillance. However, a formidable challenge remains in the automated assessment of localized geohazards(e.g., mining-induced land degradation). This difficulty stems from a fundamental limitation of existing remote sensing paradigms: their inability to perceive the concealed relational veins—the complex network of spatial, semantic, and functional linkages that connect disparate land parcels and govern their true interpretation. To overcome this challenge, we propose Venagnosis, a novel knowledge-guided graph reasoning framework that enables a paradigm shift from context-agnostic classification to context-aware relational reasoning. The Venagnosis framework is realized through three synergistic innovations. First, a Synergistic Parcel Characterization (SPC) module employs deep learning models, including a topography-focused CNN and MLP-based enhancers, to distill high-fidelity, information-dense embeddings from raw DEM, multispectral, and morphological data. Second, we construct a knowledge-guided, multi-relational graph that explicitly models the "relational veins" by encoding domain knowledge into its topology through spatial, semantic, and functional corridor edge schemas. Third, at the framework's core is the Progressive Fusion Graph Network (PF-GNN), an advanced graph attention architecture featuring adaptive input enhancement, hierarchical dynamic gating for information flow control, and a stateful cross-layer fusion mechanism for comprehensive reasoning. Comprehensive evaluations on a large-scale, national dataset of China demonstrate the state-of-the-art performance of Venagnosis. The framework achieves an F1-score of 0.8602 and an AUPRC of 0.9056, significantly outperforming competitive baselines. Rigorous ablation studies empirically validate the indispensable contribution of each architectural component. Furthermore, cross-regional generalization experiments on two distinct provincial datasets—from Henan (53,963 parcels) and Shanxi (24,833 parcels)—validate the framework's robustness and practical utility, successfully identifying unrecorded potential hazards. This work establishes a powerful, generalizable, and interpretable new paradigm for the intelligent assessment of geohazards from multi-source geospatial data.